Speaker
Description
Hierarchical quantum spectral encoders often include a global ``fusion'' stage that
entangles otherwise disjoint frequency subspaces, at the price of doubling the
variational quantum parameter count and the two-qubit gate depth. We ask a single,
practical question: \emph{does this fusion stage actually improve classification, and
does the answer change as training data grows?} We run a controlled fusion-on/off
ablation of the Hierarchical Quantum Discrete Cosine Transform (H-QDCT) encoder across
four training-set sizes ($N \in \{500,1000,2000,5000\}$, the largest only on datasets
whose native pools allow it) on six MedMNIST benchmarks ($K \in \{2,5,7,8,9\}$), with
five-to-ten seeds per cell and a paired Wilcoxon design. Pooled across $95$ paired
observations on the four multiclass tasks ($K\!\geq\!5$), the fusion stage yields no
consistent benefit (median $\Delta\!=\!-0.010$ macro-F1, $p\!=\!0.148$); across all six
datasets only $1$ of $21$ per-cell tests reaches significance, consistent with
multiple-testing noise, and increasing $N$ up to $5000$ does not turn the small effect
into a reliable one. Removing fusion halves the variational quantum parameters
($144\!\to\!72$) and the per-patch two-qubit gate count at no measurable accuracy
cost---a free simplification for NISQ deployment. We conclude that, for this class of
encoder, classical re-integration of the subspace outputs is sufficient and on-circuit
cross-band fusion is dispensable.
| Topic | Quantum variational circuits and their applications |
|---|